• No results found

Improved Methods for Detecting Acquirer Skills

N/A
N/A
Protected

Academic year: 2021

Share "Improved Methods for Detecting Acquirer Skills"

Copied!
42
0
0

Loading.... (view fulltext now)

Full text

(1)

DIVISION OF THE HUMANITIES AND SOCIAL SCIENCES

CALIFORNIA INSTITUTE OF TECHNOLOGY

PASADENA, CALIFORNIA 91125

Improved Methods for Detecting Acquirer Skills

Eric de Bodt and Jean-Gabriel Cousin

Université Lille Nord de France

Richard Roll

California Institute of Technology

SOCIAL SCIENCE WORKING PAPER #1419

May 2016

(2)

Improved  Methods  for  Detecting  Acquirer  Skills    

   

Eric  de  Bodt1,2,  Jean-­‐Gabriel  Cousin1,2,  and  Richard  Roll3  

1  Univ.  Lille  Nord  de  France,  1  place  Déliot  -­‐  BP381,  F-­‐59020  Lille,  France   2  SKEMA  Business  School,  Avenue  Willy  Brandt,  F-­‐59777  Euralille,  France  

3  California  Institute  of  Technology,  1200  East  California  Boulevard,  Pasadena,  CA  991125,  USA  

   

This  draft:  February  2016    

ABSTRACT  

 

Large  merger  and  acquisition  (M&A)  samples  feature  the  pervasive  presence  of  repetitive  acquirers.   They   offer   an   attractive   empirical   context   for   revealing   the   presence   of   acquirer   skills   (persistent   superior  performance).  But  panel  data  M&A  are  quite  heterogeneous;  just  a  few  acquirers  undertake   many  M&As.  Does  this  feature  affect  statistical  inference?  To  investigate  the  issue,  our  study  relies   on  simulations  based  on  real  data  sets.  The  results  suggest  the  existence  of  a  bias,  such  that  extant   statistical   support   for   the   presence   of   acquirer   skills   appears   compromised.   We   introduce   a   new   resampling  method  to  detect  acquirer  skills  with  attractive  statistical  properties  (size  and  power)  for   samples   of   acquirers   that   complete   at   least   five   acquisitions.   The   proposed   method   confirms   the   presence  of  acquirer  skills  but  only  for  a  marginal  fraction  of  the  acquirer  population.  This  result  is   robust   to   endogenous   attrition   and   varying   time   periods   between   successive   transactions.   Claims   according   to   which   acquirer   skills   are   a   first   order   factor   explaining   acquirer   cross-­‐sectional   cumulated  abnormal  returns  appears  overstated.  

 

JEL  classification:  G34  

Keywords:  mergers  and  acquisitions,  skills,  attrition,  panel  data    

(3)

2  

Do   some   acquirers   persistently   display   superior   performance?   This   question   is   important,   because  such  persistence  implies  the  existence  of  acquisition  skills,  achieved  through  the  acquirer’s   culture,  history,  expertise,  management  style,  or  access  to  funding  sources;  skills  that  are  difficult  for   other   firms   to   replicate.   Firms   missing   this   expertise   should   then   focus   on   internal   innovation   and   organic  growth.  

A  pervasive  feature  of  the  market  for  corporate  control  is  the  presence  of  repetitive  acquirers.   According  to  Aktas  et  al.  (2012),  in  a  sample  of  321,610  merger  and  acquisition  (M&A)  transactions   between   1992   and   2009,   approximately   25%   involved   acquirers   that   had   undertaken   at   least   five   acquisitions   during   that   period.   These   repetitive   acquirers   create   a   panel   data   structure   in   M&A   samples,  offering  a  rich  opportunity  to  test  various  theories  and  predictions.  For  example,  Schipper   and   Thompson   (1983)   and   Malatesta   and   Thompson   (1985)   investigate   investors’   anticipation   of   acquisition   programs.   Referring   to   the   hubris   hypothesis   (Roll,   1986)   and   data   that   show   that   acquirers’   cumulative   abnormal   returns   (CAR)   decline   during   acquisition   programs   (Fuller   et   al.,   2002),  several  authors  argue  that  repetitive  acquirers  develop  overconfidence  (e.g.,  Billett  and  Qian,   2008),   though   Aktas   et   al.   (2009)   question   whether   a   declining   CAR   is   unambiguous   evidence   of   hubris.  Hayward  (2002)  also  examines  the  conditions  in  which  firms  develop  acquisition  experience.    

Building  on  an  econometric  approach  designed  by  Bertrand  and  Schoar  (2003;  B&S  hereafter)  to   test   for   the   presence   of   a   particular   management   style,   some   studies   also   have   begun   addressing   acquirer   skills   (Golubov   et   al.,   2015).   This   setup   relies   on   CEO   fixed   effects   (FE),   such   that   B&S   interpret  significant  CEO  FE  as  evidence  of  a  management  style.  In  particular,  they  focus  on  changes   in  the  R-­‐square  and  adjusted  R-­‐square  values  when  switching  from  a  classical  ordinary  least  squares   (OLS)  estimator  to  the  data  panel  FE  least  squares  dummy  variable  (LSDV)  estimator,  as  well  as  on   Fisher  test  of  the  joint  significance  of  FE  (FE  Fisher  Statistic).  Yet  the  importance  they  attribute  to  the  

R-­‐square  and  adjusted  R-­‐square  values  is  puzzling.    These  statistics  are  indeed  most  often  used  as   goodness  of  fit  measures,  not  statistical  tests.  Even  if  asymptotic  distributions  exist,  these  depend  on   unknown   parameters   (see   Ohtani,   2000).   Moreover,   without   a   clear   null   hypothesis,   the   interpretation  of  their  results  is  ambiguous.  It  is  worthwhile  also  to  note  that  statistical  findings  using   the  R-­‐square  or  the  adjusted  R-­‐square  offer  no  insights  into  the  sign  associated  with  skills  (i.e.,  under-­‐   or   over-­‐performance).   The   FE   Fisher   test   has   the   potential   to   reject   the   clearly   defined   null   hypothesis  of  no  acquirer  skills  though.    

In   the   specific   case   of   M&A   sample   the   data   sets   are   panel   data   sets   strongly   unbalanced,   characterized  by  attrition  as  defined  in  Wooldridge  (2002).  For  example,  in  Aktas  et  al.’s  (2012)  very   large  sample,  more  than  50%  of  the  transactions  involve  acquirers  that  make  only  one  acquisition.   The  CAR  cross-­‐sectional  variation  of  one-­‐time  acquirers  then  can  be  captured  fully  and  mechanically   by  an  FE  estimator.  But  does  this  attrition  pattern  affect  the  R-­‐square,  adjusted  R-­‐square,  and  Fisher  

(4)

3  

test?   To   what   extent   does   it   contaminate   empirical   findings?   In   addition   to   answering   these   questions,  we  seek  an  alternative  testing  procedure  for  detecting  acquirer  skills  that  might  be  more   robust  to  the  specific  attrition  pattern.    

To   do   so,   our   analyses   use   a   sample   of   12,707   transactions   completed   during   1990–2011   by   4,507   unique   acquirers.   Our   sample   selection   criteria   match   those   of   Golubov   et   al.   (2015):   domestically   controlled   transactions,   public   acquirers,   targets   of   all   statuses   (public,   private,   subsidiaries),  completed  transactions,  deal  value  of  at  least  US  $1  million  as  reported  in  the  Thomson   Securities   Data   Company   (SDC)   database,   relative   transaction   size   at   least   equal   to   1%,   and   no   financial   industries   (standard   industrial   classification   [SIC]   codes   6000–6999).   Our   sample   includes   27.11%  fully  cash-­‐paid  deals  and  15.99%  public  targets.  The  average  deal  value  is  US$  377  million,   and  the  average  acquirer  CAR  is  1.71%.  These  statistics  are  all  consistent  with  previous  reports  on   similar  sample  types  (e.g.,  Moeller  et  al.,  2004).  Of  the  4,507  unique  acquirers,  1,859  are  one-­‐time   acquirers   (41.25%),   whereas   781   (17.33%)   engaged   in   at   least   five   transactions   during   1990–2011.   These  two  figures  highlight  the  strong  attrition  in  this  typical  M&A  data  set.  

To   assess   the   influence   of   attrition   on   inferences   based   on   the   B&S   approach,   we   conducted   simulation  studies,  in  the  style  of  Brown  and  Warner  (1985;  B&W  hereafter),  using  our  M&A  sample   and   adding   simulated   acquirer   skills.   We   manipulate   the   attrition   pattern   of   the   generated   M&A   samples,  that  is,  the  percentage  of  acquirers  that  complete  a  particular  number  of  transactions.  For   example,  we  simulate  samples  in  which  63.21%  of  acquirers  are  one-­‐time  acquirers,  23.26%  are  two-­‐ time   acquirers,   8.56%   are   three-­‐time   acquirers,   and   so   on.   In   seven   attrition   patterns   (Figure   1),   attrition  in  the  number  of  transactions  shifts  from  a  rapid  pace  (right-­‐skewed  attrition),  as  typically   observed  for  M&A  samples,  to  a  slow  pace  (left-­‐skewed  attrition).  Thus  we  can  examine  the  impact   of   panel   attrition   on   the   ability   to   detect   acquirer   skills.   Each   M&A   transaction   also   is   assigned   a   random  acquirer  skill  level,  drawn  from  a  zero  mean  Gaussian  distribution  with  a  given  variance.  This   skill   applies   to   all   transactions   by   a   given   acquirer   (i.e.,   perfectly   persistent).   The   variance   of   the   Gaussian  distribution  then  drives  the  importance  and  heterogeneity  of  skills  in  the  acquirer  sample.   We  regress  the  acquirer  CAR  on  acquirer  fixed  effects  (FE)  and  the  set  of  control  variables  suggested   by  Golubov  et  al.  (2015),  then  analyze  the  behavior  of  the  FE  Fisher  test.  We  report  also  R-­‐square  and   adjusted  R-­‐square   values,   to   parallel   existing   literature.   We   repeat   this   process   1,000   times,   with   different  combinations  of  the  variances  of  abnormal  returns  and  attrition  patterns.  In  turn,  we  offer   three  key  insights.  

First,   in   the   absence   of   simulated   acquirer   skills,   when   switching   from   the   OLS   to   the   LSDV   estimator,  the  R-­‐square  value  increases  dramatically  in  case  of  right-­‐skewed  attrition.  With  the  LSDV   estimator,   the  R-­‐square   is   only   weakly   reactive   to   the   importance   of   simulated   skills   with   right-­‐ skewed   attrition.   We   conclude   that   the   observation   of   an   increase   in   the  R-­‐square   value   when  

(5)

4  

switching  from  OLS  to  LSDV  cannot  reveal  insights  into  the  presence  or  absence  of  acquirer  skills  in   typical  M&A  samples.  

Second,  again  in  the  absence  of  simulated  acquirer  skills,  the  increase  in  the  adjusted  R-­‐square   value   that   results   from   the   switch   from   the   OLS   to   the   LSDV   estimator   is   limited.   With   the   LSDV   estimator,   the   adjusted  R-­‐square   is   moreover   more   reactive   to   the   importance   of   simulated   skills,   independently  of  the  attrition  pattern.  Thus,  it  is  more  suited  to  detect  acquirer  skills  than  R-­‐square   is.  But,  as  mentioned  here  above,  the  adjusted  R-­‐square  is  essentially  goodness  of  fit  measure,  with   an  asymptotic  distribution  depending  on  unknown  parameters,  that  offer  therefore  a  limited  route   to   test   statistical   evidence   of   the   presence   of   acquirer   skills.   It   also   is   silent   about   the   sign   of   the   detected  skills.  

Third,   the   FE   Fisher   Statistic,   similar   to   the   adjusted  R-­‐square,   displays   reactivity   to   the   importance   of   simulated   skills.   Yet   our   B&W   simulations   highlight   that   the   FE   Fisher   Statistic   size   depends  on  the  attrition  pattern.  In  a  case  of  right-­‐skewed  attrition  (as  typically  observed  in  M&A   samples),  the  FE  Fisher  Statistic  is  vastly  over-­‐sized:  in  the  absence  of  simulated  acquirer  skills,  the   null  hypothesis  of  no  acquirer  skills  is  rejected  far  too  often,  according  to  the  chosen  confidence  level   (type   I   error).   Therefore,   the   use   of   the   FE   Fisher   test   to   detect   acquirer   skills   leads   to   potentially   strongly  biased  inferences.  

Reflecting  these  findings  regarding  the  FE  Fisher  Statistic  size  issue  and  the  extent  to  which  FE   estimation   precision   depends   on   the   number   of   acquisitions   by   the   acquirer,   we   propose   a   new   resampling-­‐based   method   to   detect   acquirer   skills   and   designated   as   RBSD   for  Resampling   Based   Method  for  Skills  Detection.  It  builds  on  a  simple  idea:  reconstruct  balanced  panels  for  each  number   of   acquisitions   by   an   acquirer.   By   construction,   the   generated   M&A   samples   display   no   more   attrition.  We  then  analyze  the  size  and  power  of  RBSD  using  the  set  of  B&W  simulations  adopted  for   the   B&S   procedure.   Here   again,   some   clear   conclusions   emerge:   the   FE   Fisher   Statistic   is   correctly   sized,   even   if   the   sample   displays   right-­‐skewed   attrition.   In   the   power   analysis   (i.e.,   ability   of   the   RBSD  FE  Fisher  Statistic  to  reject  the  absence  of  acquirer  skills  in  presence  of  simulated  skills),  we   observe   that   power   increases   with   the   number   of   acquisitions   by   acquirer,   which   is   as   expected,   because   skills   by   definition   are   based   on   persistence.   Power   also   is   increasing   in   the   level   of   simulated  skills.  Therefore,  the  RBSD  FE  Fisher  Statistic  appears  to  be  a  valid  statistical  test  for  the   presence  of  acquirer  skills  in  a  real-­‐world  M&A  sample.    

Applying   the   RBSD,   we   finally   test   for   the   presence   of   acquirer   skills   in   our   M&A   sample.   It   confirms   the   presence   of   acquirer   FE:   At   a   10%   confidence   level,   for   balanced   samples   of   5   acquisitions   per   acquirer,   in   88.20%   of   the   generated   samples,   we   can   reject   the   absence   of   significant  acquirer  FE.  At  5%  and  1%  confidence  levels,  the  corresponding  percentages  are  73.50%   and   36.10%.   But   the   percentages   of   acquirers   displaying   statistically   significant   FE   are   low.   For  

(6)

5  

balanced  samples  of  5  acquisitions  per  acquirer,  we  find  that  6.23%,  3.06%,  and  0.65%  of  acquirer  FE   are   significant   at   10%,   5%,   and   1%   confidence   levels,   respectively.   These   results   are   robust   to   endogenous   attrition   due   to   past   performance   and   controls   for   the   varying   time   periods   between   successive   transactions.   Thus,   we   assert   that   the   claims   presenting   acquirer   skills   as   a   first-­‐order   factor  to  explain  the  cross-­‐section  of  acquirer  CAR  are  overstated.    

Our   results   accordingly   contribute   to   M&A   literature.   They   put   into   question   Golubov   et   al.’s   (2015,   p.   315)   general   conclusions   that   “acquirer   returns   are,   indeed,   best   explained   by   an   unobserved,   time-­‐invariant,   firm-­‐specific   factor.”   Their   conclusions   rely   on   the   B&S   setup   and   the   persistence  of  acquirer  performance  throughout  acquisition  programs.  We  show  that  FE  Fisher  tests   for  data  panels  that  display  strong  right-­‐skewed  attrition  are  over-­‐sized  and  can  lead  to  inaccurate   inferences1.  Golubov  et  al.  (2015)  also  report  the  presence  of  persistence  in  acquirer  performance  

through  acquisition  programs—notable,  but  not  enough  to  validate  the  presence  of  skills.  As  Aktas  et   al.  (2009)  show,  acquirer  performance  persistence  also  might  be  consistent  with  learning.    

Our   results   instead   support   Fee   et   al.’s   (2013)   challenges   to   B&S’s   results,   in   which   they   used   evidence  from  exogenous  CEO  departures  to  test  whether  these  shocks  affect  firm  behavior.  They   find   no   such   effect,   in   contrast   with   predictions   based   on   the   management   style   hypothesis.   They   also  assess  the  power  of  the  B&S  approach  for  uncovering  management  style,  scrambling  their  data   in  such  a  way  that,  by  construction,  a  management  style  effect  cannot  exit.  To  test  for  the  presence   of  a  management  style  on  this  simulated  data  set,  they  use  the  B&S  setup.  The  FE  Fisher  test  in  that   case  led  to  a  spurious  conclusion  about  the  presence  of  a  management  style  effect.  We  extend  this   analysis  to  of  the  case  of  data  panel  attrition  patterns  that  characterize  M&A  samples.  We  also  offer   the  RBSD  approach  as  an  improved  method  to  detect  acquirer  skills.  

  1. Data    

 

1.1.  M&A  Sample  

We  collect  M&A  transactions  from  the  SDC  database  over  the  1990–2011  period,  with  the  same   selection  criteria  used  by  Golubov  et  al.  (2015):    

-­‐ Domestic  transactions  (U.S.  acquirers  and  U.S.  targets);    

-­‐ Completed  control  transactions  (acquirer  holds  less  than  50%  of  the  target  shares  before  the   announcement  and  ends  up  with  100%  of  the  target  shares);  

                                                                                                                         

1  Consistent  with  our  simulation  results,  Golubov  et  al.  (2015)  report  that  for  a  subsample  of  

acquirers  that  completed  at  least  two  deals  during  1990–2011,  the  Fisher  joint  test  of  FE  significance   drops  sharply  (in  Table  2,  from  1.692  in  Panel  A  to  1.287  and  1.261  in  Panels  B  and  C).  Table  2  also   reports  that  the  adjusted  R-­‐square  value  drops  from  23.1%  in  the  full  sample  to  12%  and  6.8%,   respectively,  for  subsamples  of  serial  acquirers.  These  results  are  more  consistent  with  ours.  

(7)

6  

-­‐ Public  acquirers  and  targets  of  all  statuses  (private,  public,  subsidiaries);   -­‐ Deal  value  of  at  least  US$  1  million;  

-­‐ Relative  transaction  size  (ratio  of  the  deal  value  to  the  acquirer  market  value)  of  at  least  1%;   -­‐ Financial  industries  (SIC  codes  6000–6999)  excluded;  and    

-­‐ Necessary   information   available   in   the   CRSP   and   COMPUSTAT   databases   to   compute   the   acquirer  CAR  and  the  set  of  control  variables.  

Applying  these  criteria,  we  collected  12,707  deals.  Golubov  et  al.  (2015)  obtain  12,491  transactions   over  the  same  period.  Table  1,  Panel  A,  reports  the  number  of  deals  by  year;  Panel  B  reports  them  by   deal  order  number  (DON),  which  is  the  deal  number  in  the  sequence  of  transactions  completed  by  a   given  acquirer.  In  Table  1,  Panel  A,  the  M&A  waves  at  the  end  of  the  1990s  and  the  mid-­‐2000s  are   apparent   (see   also   Betton   et   al.,   2008).   Furthermore,   Figure   1,   Panel   B,   displays   the   well-­‐known   stylized  facts  about  the  presence  of  repetitive  acquirers,  such  that  there  are  many  one-­‐time  acquirers   (41.25%   of   all   acquirers   in   our   sample,   or   1,859   out   of   4,507),   as   well   as   some   active   repetitive   acquirers  (781  firms  completed  at  least  five  deals,  or  17.33%  of  the  sample).  These  statistics  coincide   with  previous  reports  (e.g.,  Aktas  et  al.,  2012).  

 

1.2.  Dependent  Variable    

  The  acquirer  CAR  is  the  dependent  variable.  We  calculate  it  over  a  three-­‐day  event  window   centered  on  the  deal  announcement,  as  reported  in  the  SDC  database.  We  obtain  abnormal  returns   using  the  market  model.  We  choose  an  estimation  window  from  day  –300  to  day  –91  relative  to  the   announcement  on  day  0.  Table  1  displays  the  descriptive  statistics  of  interest.  The  acquirer  average   CAR   is   1.71%,   a   figure   typical   of   large   M&A   samples   that   include   public   and   private   targets   (e.g.,   Moeller  et  al.  (2004)  report  1.10%  in  a  sample  of  12,023  transactions  during  1980–2001;  Betton  et  al.   (2008)  report  0.73%  for  a  sample  of  9,298  transactions  over  1980–2005).  As  Table  1,  Panel  A,  shows,   M&A   transactions   were   more   profitable   for   acquirers   during   the   early   1980s   (cf.   1980).   Panel   B   reveals   the   clearly   declining   trend   of   acquirer   CAR   as   a   function   of   the   DON,   which   some   authors   interpret  as  a  signal  of  hubris  or  overconfidence  (Billet  and  Qian,  2008),  though  Aktas  et  al.  (2009)   argue  that  declining  CAR  through  acquisition  programs  is  not  such  an  unambiguous  indicator.  

 

1.3.  Control  Variables    

  We   collect   a   set   of   control   variables   widely   used   in   M&A   literature   (Moeller   et   al.,   2004;   Golubov  et  al.,  2015):  

-­‐ Bidder   size:   the   bidder’s   market   value   at   the   end   of   the   fiscal   year   before   the   acquisition   announcement  in  millions  of  U.S.  dollars;  

(8)

7  

-­‐ Stock:  a  dummy  variable  equal  to  1  if  the  transaction  is  fully  paid  in  stock;   -­‐ Private:  a  dummy  variable  equal  to  1  if  the  target  is  a  private  company;   -­‐ Public:  a  dummy  variable  equal  to  1  if  the  target  is  a  public  company;   -­‐ Subsidiary:  a  dummy  variable  equal  to  1  if  the  target  is  a  subsidiary;  

-­‐ Tobin’s   Q:   the   acquirer   market   value   of   assets   (defined   as   the   book   value   of   total   assets   minus  common  equity  plus  the  market  value  of  equity)  divided  by  the  acquirer  book  value  of   assets;  

-­‐ Run-­‐up:  the  market-­‐adjusted  buy  and  hold  return  of  the  acquirer’s  stock  price  from  day  –210   to  day  –11  with  respect  to  the  announcement  date;  

-­‐ FCF   (free   cash-­‐flow):   the   acquirer’s   operating   income   before   depreciation   minus   interest   expense  and  income  taxes  plus  changes  in  deferred  taxes  and  investment  tax  credit  minus   dividends  on  both  preferred  and  common  share  divided  by  the  book  value  of  total  assets;   -­‐ Leverage:   the   acquirer’s   long-­‐term   debt   divided   by   the   market   value   of   assets,   defined   as  

above;  

-­‐ Sigma:  the  standard  deviation  of  the  acquirer  market-­‐adjusted  daily  returns  from  day  –210  to   day  –11  with  respect  to  the  announcement  date;  

-­‐ Relative  size:  the  ratio  of  the  deal  value  to  the  acquirer  market  value;  

-­‐ Relatedness:  a  dummy  variable  equal  to  1  if  the  bidder  and  the  target  operate  in  the  same   industry  at  the  two-­‐digit  SIC  code  level;  

-­‐ Tender  offer:  a  dummy  variable  equal  to  1  if  the  deal  is  classified  as  a  tender  offer  in  the  SDC   database;  and    

-­‐ Hostile:  a  dummy  equal  to  1  if  the  transaction  is  classified  as  hostile  in  the  SDC  database.   The   acquirer   market   value   and   acquirer   financial   statements   items   are   collected   at   the   end   of   the   fiscal  year  before  the  M&A  announcement  date.  

  Using  the  descriptive  statistics  by  year  and  by  DON  in  Table  1,  we  can  compare  our  data  with   Moeller  et  al.’s  (2004),  though  their  sample  covers  a  different  period  (1980–2001).2  Our  average  deal  

value  is  US$  377  million,  versus  US$  257  million  in  Moeller  et  al.  (2004),  which  matches  the  secular   increase  in  deal  values.  A  corresponding  increase  appears  in  the  average  acquirer’s  market  value— US$  2,777  million  versus  US$  1,708  million.  The  percentage  of  fully  cash  paid  transactions  is  27.11%   in   our   sample   versus   40.44%   in   Moeller   et   al.’s   (2004),   and   the   acquirer   Tobin’s   Q   is   2.19   in   our   sample   versus   1.89.   According   to   these   statistics,   our   M&A   sample   does   not   display   unexpected                                                                                                                            

2  A  comparison  with  Golubov  et  al.  (2015),  who  analyze  the  same  period,  is  not  possible  because   they  do  not  report  descriptive  statistics.  

(9)

8  

features.   Finally,  several  of  the  control  variables  display  a  clear  time  trend,  including  Cash  dummy   (strongly  increasing),  Public  dummy  (decreasing),  and  Relative  size  (strongly  decreasing).  

  According  to  the  statistics  by  DON  in  Table  1,  Panel  B  (as  also  reported  by  Aktas  et  al.,  2012),   several  variables  increase  with  the  DON,  including  deal  values  (suggesting  that  acquirers  start  with   small   transactions),3  acquirer   market   value   (i.e.,   repetitive   acquirers   are   larger   firms   or   else   grow  

through  acquisition),  Cash  dummy,  Leverage,  and  Tender  offers.  In  contrast,  Tobin’s  Q,  Sigma,  and   Relative  size  decrease,  consistent  with  the  increase  in  acquirer  size  in  the  latter  case.    

  Table  2  contains  the  results  of  the  multivariate  analyses  of  acquirer  CAR.  Column  1  reports   estimates  obtained  with  the  classical  OLS  estimator,  and  then  in  column  2,  we  add  year  FE,  and  in   column  3,  we  present  results  obtained  with  the  acquirer  FE  estimator.  The  comparison  of  the  column   2   results   with   Golubov   et   al.’s   (2015)   table   1,   using   the   same   estimator,   reveals   nearly   the   same   results   (though   the   interactions   of  Public   and  Cash   and   of  Private  and  Stock  are   significant   in   our   case).  As  we  show  in  column  3,  the  acquirer  FE  estimator  is  relevant  for  panel  data,  and  bidder  size   negatively  affects  acquirer  CAR,  as  does  Run-­‐up,  Leverage,  and  the  interaction  between  acquiring  a   public  target  and  paying  in  stock.  Sigma  (i.e.,  acquirer  stock  return  standard  deviation),  Relative  size,   and   the   interaction   between   acquiring   a   private   company   and   paying   in   stock   all   have   positive   impacts   on   acquirer   CAR.   Comparisons   of   the   results   across   columns   show   that   the   significance   of   Bidder  size,  Run-­‐up,  Sigma,  the  interaction  of  Public  target  and  Stock  acquisition,  and  the  interaction   of  Private  and  Stock  acquisition  are  all  robust  to  the  chosen  estimator.  In  addition,  Leverage  changes   sign   (from   positive   to   negative)   when   adopting   the   acquirer   FE   estimator.   The   positive   relation   between   acquirer   CAR   and   acquirer   leverage   is   a   cross-­‐sectional   phenomenon   (more   leveraged   acquirers  complete  more  value-­‐creating  transactions  on  average),  not  a  time-­‐series  one  (increased   leverage  for  a  given  acquirer  leads  to  a  decrease  in  CAR  on  average).    

 

2. Econometric  Estimators  and  Statistics  of  Interest  

The  B&S  approach  relies  on  a  fixed  effects  (FE)  panel  data  regression.  The  population  regression   model  takes  the  following  form  for  the  present  case:  

 

𝑦!,! =𝛼+𝒙!,!𝜷+𝑣!+𝜀!,!,           (1)  

 

                                                                                                                         

3  We  note  one  extremely  large  transaction,  at  DON  eight:  AOL-­‐Time  Warner,  with  a  deal  size  of  US$   168  billion.  

(10)

9  

where  𝑖  is   the   acquirer   index,  𝑡  is   the   transaction   index,  𝑦!,!  is   the   variable   of   interest   (eg.,   a   firm   performance  measure  such  as  the  Return  On  Assets),    𝒙!,!  is  a  vector  of  explanatory  variables,  𝜷  is   the   corresponding   vector   of   coefficients,  𝑣!+𝜀!,!  is   the   error   term,  𝑣!  is   acquirer-­‐specific   error   capturing  time-­‐constant  unobservable  factors,  and  𝜀!,!  is  the  “classic”  error  term  (uncorrelated  with   𝒙!,!  and  𝑣!).   B&S   use   the   LSDV   regression   to   estimate   firm   FEs,   which   essentially   estimates   the  

following  regression  model:    

𝑦!,! =𝑐+𝛼!+𝒙!,!𝜷+𝜀!,!,             (2)  

  where  𝛼!  is   a   dummy   variables   equal   to   1   for   acquirer   i.   The   OLS   estimates  𝛼!  are   unbiased  

estimators  of  𝑣!,4  the  firm  FE  of  interest.5    

Next,  B&S  study  the  behavior  of  R-­‐square  and  adjusted  R-­‐square,  two  statistics  classically  used  as   goodness   of   fit   measures,   and   of   the   Fisher   joint   significance   test   of   FE,   computed   using   LSDV   estimates:       𝑅! !"#$ =1−!!"!""!"#$  ,               (3)   𝐴𝑑𝑗  𝑅! !"#$=1− !!! !!!!!×(1−𝑅!!"#$)  ,  and         (4)   𝐹= 𝑹  𝜶!𝒓!𝑹  𝚺𝜶𝑹!!!𝑹  𝜶!𝒓 !  ,             (5)     where  𝑅!

!"#$  refers  to  the  LSDV  R-­‐square,  𝐴𝑑𝑗  𝑅!!"#$  is  the  adjusted  LSDV  R-­‐square,  𝑆𝑆𝑅!"#$  is  the   sum  of  squared  LSDV  residuals,  𝑇𝑇𝑆  is  the  total  sum  of  squares,  𝑛  is  the  number  of  observations  in   the   sample,  𝑘  is   the   number   of   estimated   coefficients,  𝐹  is   the   FE   Fisher   Statistic,6  (𝑹  𝜶𝒓)  is   the  

matrix  of  linear  restrictions  (all  FE  =  0);  𝑹  is  the  matrix  of  linear  restriction  coefficients;  𝜶  is  the  vector                                                                                                                            

4  The  𝛼!  OLS  estimates  are  not  consistent  because,  for  a  given  𝑇  (number  of  transactions),  their   asymptotic  variance  does  not  converge  to  zero  as  𝑁  (the  number  of  acquirers)  goes  to  infinity  (see   Greene,  2011).  

5  Other  estimators  are  available  to  estimate  the  vector  𝜷  of  coefficients.  The  pooled  estimator  runs  a   classical  regression,  ignoring  the  presence  of  time-­‐constant  unobservable  factors;  the  within  

estimator  runs  a  regression  on  group-­‐demeaned  observations;  and  the  between  estimator  runs  on   the  variation  of  group  means  around  the  overall  mean.  None  of  these  estimators  provides  estimates   of  the  𝛼!  of  interest  though.  

6  Even  if  the  OLS  𝛼!  estimates  are  not  consistent,  the  FE  Fisher  is  valid  to  test  the  null  hypothesis  of  

(11)

10  

of  estimated  FE;  𝒓  is  a  vector  of  constants  (0  in  our  case);  𝚺𝜶  is  the  estimated  variance–covariance   matrix  of  𝜶;  and  𝑞  is  the  number  of  restrictions.  

Even   if   asymptotic   distributions   of   the  R-­‐square   and   adjusted  R-­‐square   statistics  exist,   these   depend  on  unknown  parameters  (see  Ohtani,  2000).  Moreover,  without  a  clear  null  hypothesis,  their   interpretation   remains   ambiguous.   These   are   most   probability   the   reasons   explaining   their   use   restricted  to  goodness  of  fit  measures.  On  the  contrary,  the  FE  Fisher  statistic  is  classically  used  for   formal   statistical   inferences   because   it’s   asymptotic   distribution   depends   only   on   parameters   that   can  be  estimated  using  their  sample  counterparts.  It  is  also  important  to  recognize  that  the  FE  Fisher  

tests   the   null   hypothesis   that   “all   FEs   jointly   equal   0,”   which   is   rejected   if   even   only   one   of   the   constraint  is  rejected  while  all  other  ones  are  satisfied.  Although  it  is  informative  about  the  presence   of  at  least  one  skilled  acquirer  in  our  case,  the  FE  Fisher  Statistic  provides  no  information  about  the   frequency  of  the  phenomenon  (i.e.,  number  of  skilled  acquirers  in  the  sample).  Moreover,  the  Fisher   test   is   bilateral   and   reacts   to   the   presence   of   both   positively   and   negatively   significant   FE.   It   does   therefore  not  discriminate  between  positive  (value-­‐creating)  and  negative  (value-­‐destroying)  skills.  

 

3. Simulation  Procedures    

3.1.  Attrition  Pattern  

  At   the   heart   of   our   study   is   the   simulation   of   different   attrition   patterns,   depicting   the   percentage  of  acquirers  that  complete  a  given  number  of  deals  (we  limit  ourselves  to  a  maximum  of   10   deals,   because   fewer   than   3.6%   of   the   transactions   have   a   DON   above   10   during   our   sample   period7).  We  simulate  both  a  rapid  pace  of  attrition  (right-­‐skewed  attrition)  and  slow  pace  of  attrition  

(left-­‐skewed  attrition)  using  Equation  6:     %  𝑆𝑎𝑚𝑝𝑙𝑒!"=100  × !!  ×!" !!  ×! ! !!!  ,         (6)     where  %  𝑆𝑎𝑚𝑝𝑙𝑒!"  is   the   percentage   of   acquirers   having   completed  𝑁𝐷  deals   in   the   simulated   sample   and  𝑁  is   the   maximum   number   of   deals   completed   by   any   given   acquirer   in   the   simulated   sample  (10  in  the  present  case).  We  choose  𝛼  equal  to  –1,  –0.5,  and  –0.1  for  rapid  attrition  (right-­‐ skewed  attrition)  and  to  0.1,  0.5,  and  1  for  slow  attrition  (left-­‐skewed  attrition).    

                                                                                                                         

7  Generating  simulated  samples  containing  an  important  proportion  of  highly  repetitive  acquirers   leads  indeed  to  incorporate  many  times  the  same  acquirers  in  each  simulated  samples.  

(12)

11  

We   also   add   the   constant   repartition   of   transactions   by   numbers   of   deals   (10%   of   the   sample).  Figure  1,  Panel   A,  presents  the   different  attrition  patterns   obtained   using   this   procedure.   From   left   to   right,   we   shift   from   a   drastically   right-­‐skewed   attrition   pattern   to   a   drastically   left-­‐ skewed   attrition   pattern.   The   first   three   columns   correspond   to  𝛼  equal   to   –1,   –0.5,   and   –0.1;   the   middle   column   indicates   the   constant   repartition;   and   the   last   three   columns   reflect   the   corresponding  left-­‐skewed  attrition  patterns.  Panel  A  also  displays  the  corresponding  percentages  of   transactions   included   in   the   sample   for   each   number   of   deals.   Figure   1,   Panel   B,   displays   the   percentage  of  transactions  by  number  of  deals  in  our  actual  M&A  sample.  A  comparison  of  Panels  A   and  B  reveals  that  the  attrition  pattern  featuring  our  M&A  sample  closely  corresponds  to  the  right-­‐ skewed  attrition  pattern  in  which  𝛼  equals  –0.5.  

 

3.2.  Brown  and  Warner  (B&W,  1985)  Simulations  

  We  study  the  interactions  between  the  behavior  of  the  R-­‐square,  the  adjusted  R-­‐square,  the   FE  Fisher  Statistic,  and  attrition  patterns  by  implementing  a  B&W-­‐style  approach.  We  refer  to  B&W   because   the   simulation   environment   relies   on   a   real   data   set—namely,   our   M&A   sample—not   a   simulated   one   (such   that   we   would   have   implemented   a   Monte   Carlo   approach).   The   simulation   procedure  is  as  follows:  

(i) Begin  with  the  actual  M&A  sample  (Section  1).  We  limit  the  sample  to  transactions  with   DON  less  than  or  equal  to  10  to  match  the  simulated  attrition  patterns,  leaving  a  sample   of  12,253  transactions.    

(ii) Randomly  assign  M&A  transactions  to  acquirers  by  shuffling  acquirer  PERMNOs  (i.e.,  the   CRSP   database   permanent   number,   which   is   unique   to   each   firm).   We   thus   create   an   M&A  sample  under  a  null  hypothesis  of  no  skills  (we  break  any  systematic  relationship   between  a  given  acquirer  and  given  M&A  transactions).  

(iii) Model   skills   as   a   random   drawing   in   a   Gaussian   distribution   of   abnormal   returns   (denoted  𝑁(0,𝜎!")).  

(iv) Select  one  attrition  pattern  and  randomly  draw  1,000  sub-­‐samples  of  500  deals  so  that   the  attrition  pattern  is  respected.8  

(v) For  each  subsample:  

a. For  each  acquirer  𝑖  in  the  subsample:   1.Draw  a  skill  𝐴𝑅!,!"  in  𝑁(0,𝜎!");  

                                                                                                                         

8  In  robustness  checks,  we  replicate  this  exercise  with  a  constant  number  of  acquirers  and  a  constant   number  of  degrees  of  freedom  (see  Appendices  1  and  2).  

(13)

12  

2.Add  𝐴𝑅!,!"  to  the  acquirer  CAR  of  all  transactions  randomly  attributed  to  the   acquirer  in  step  (ii).  

b. Estimate  the  same  acquirer  CAR  regression  as  reported  in  Table  2,  column  3,  using   the  LSDV  FE  estimator;  

c. Collect   the  R-­‐square,   adjusted  R-­‐square,   and   FE   Fisher   Statistic   obtained   in   the   previous  step.  

(vi) Compute   the   average   R-­‐square,   adjusted   R-­‐square,   and   FE   Fisher   Statistic   values   obtained  over  the  1,000  subsamples.  

We   execute   this   procedure   for   the   combination   of   seven   attrition   patterns   and   for   values   of  𝜎!"   ranging  from  0%  to  5%  in  steps  of  1  pp.  The  resulting  42  combinations  allow  us  to  analyze  in  depth   the   interactions   among   attrition   patterns   in   M&A   samples,   acquirer   skills,   and   summary   statistics   obtained  using  the  LSDV  FE  estimator.  The  selected  values  of  𝜎!"  are  such  that  the  no  simulated  skills   case   is   taken   into   account   𝜎!" =0 ,  and   simulated   skills   are   on   an   order   of   magnitude   of   the   average  acquirer  CAR  reported  in  prior  literature.  The  higher  𝜎!",  the  greater  the  probability  that  a   given  acquirer  will  be  imputed  a  high  𝐴𝑅!,!"  (in  absolute  value)  for  each  of  its  acquisitions,  such  that   more  skilled  acquirers  will  be  present  in  the  generated  sample.  

 

4. Sample  Attrition,  LSDV  R-­‐square,  Adjusted  R-­‐square,  and  FE  Fisher  Test  

Table   3   and   Figure   2   summarize   our   B&W   simulation   results.   Table   3   comprises   three   panels,   focused  on  the  behavior  of  the  R-­‐square  (Panel  A),  the  adjusted  R-­‐square  (Panel  B),  and  the  Fisher   joint  significance  test  of  acquirer  FE  (Panel  C).  In  each  panel,  the  first  column  reports  the  level  of  𝜎!"   used   to   simulate   acquirer   skills.   The   first   two   rows   refer   to   the   case   of   no   acquirer   skills,   and   the   difference   between   the   two   rows   reflects   the   econometric   estimator   used   (except   in   Panel   C,   because  we  cannot  compute  a  Fisher  FE  test  with  the  OLS  estimator).  The  second  column  specifies   the   econometric   estimator:   OLS   pooled   regression  or   the   LSDV   regression   model   (see   Equation   2).   Therefore,  row  1  is  the  benchmark  case,  row  2  highlights  the  consequences  of  switching  from  OLS  to   LSDV  in  the  case  of  no  acquirer  skills,  and  rows  3–7  explore  the  consequences  of  an  increase  in  𝜎!"   used  to  simulate  acquirer  skills.  Columns  3–9  correspond  to  the  seven  attrition  patterns  introduced   in  Section  2  and  presented  in  Figure  1,  from  right-­‐  to  left-­‐skewed  attrition  patterns.  Figure  2  displays   the  evolution  of  the  three  statistics  of  interest  (R-­‐square,  adjusted  R-­‐square,  and  Fisher  test)  along   the  seven  attrition  patterns  (reproduced  in  the  overlay),  in  the  case  of  no  acquirer  skills  (Panel  A)  and   when  𝜎!"  equals  5%  (Panel  B).    

  We  comment  first  on  the  R-­‐square  results  (Table  3,  Panel  A).  In  the  case  of  no  acquirer  skills   (rows   1   and   2),   switching   from   OLS   to   LSDV   dramatically   increases   the  R-­‐square   in   the   case   of   attrition   pattern   1   (most   right-­‐skewed   attrition).   The  R-­‐square   goes   from   14.66%   to   74.44%,   a  

(14)

13  

fivefold  increase.  In  the  case  of  attrition  pattern  7  (most  left-­‐skewed  attrition),  the  increase  is  still   impressive   (from   14.77%   to   24.21%)   but   significantly   lower.   These   results   exactly   match   our   expectations,   because   acquirer   FE   capture   all   acquirer   cross-­‐sectional   variation   for   one-­‐time   acquirers.  Thus,    

-­‐ in   the   case   of   a   right-­‐skewed   attrition   pattern,   the   M&A   sample   is   characterized   by   the   presence   of   many   one-­‐time   acquirers,   for   which   FE   captures   100%   of   the   cross-­‐sectional   variation   of   acquirer   CAR   (one   constant   for   each   one-­‐time   acquirer).   An   increase   in   the  R-­‐ square  from  the  OLS  to  the  LSDV  estimation  thus  is  no  evidence  of  the  presence  of  acquirer   skills;  

-­‐ in  the  case  of  a  left-­‐skewed  attrition  pattern,  the  M&A  sample  incorporates  many  repetitive   acquirers  (63.21%  of  acquirers  are  ten-­‐time  acquirers).  For  these  acquirers,  the  acquirer  FE   does  not   capture   the   time   variation   of   acquirer   CAR.   The   more   they   are   present   in   the   sample,  the  lower  is  the  increase  in  the  explained  variance  due  the  presence  of  acquirer  FE.   These  results  highlight  an  important  phenomenon:  attrition  patterns  drastically  affect  the  behavior   of  the  R-­‐square,  independent  of  the  presence  of  acquirer  skills,  when  switching  from  OLS  to  LSDV.   This  R-­‐square  behavior  is  clearly  apparent  in  Figure  2,  Panel  A.  

  The  second  striking  pattern  of  behavior  for  the  R-­‐square  is  that,  using  the  LSDV  estimator,  it   is   almost   insensitive   to   simulated   acquirer   skills   for   right-­‐skewed   attrition   patterns   (it   oscillates   around  75%).  Only  for  the  left-­‐skewed  attrition  pattern  (many  repetitive  acquirers)  does  the  R-­‐square   become   more   sensitive   to   simulated   skills,   ranging   from   25%   for   low   levels   of  𝜎!"  to   37%   for   the   highest  level.  This  result  again  highlights  that  the  LSDV  R-­‐square  is  silent  about  the  presence  acquirer   skills  in  a  situation  with  a  right-­‐skewed  attrition  pattern.    

Regarding  the  adjusted  R-­‐square,  Table  3,  Panel  B,  reports  fundamentally  different  results.  That   is,  in  the  case  of  no  acquirer  skills,  when  switching  from  OLS  to  LSDV,  the  adjusted  R-­‐square  moves   from  7.83   %  to  9.08%    for  attrition  pattern  1.  This  result  is  to  be  expected;  the  adjusted  R-­‐square   explicitly  accounts  for  the  number  of  estimated  parameters  (see  Equation  4).  But  is  the  acquirer  R-­‐ square  better  able  to  detect  the  presence  of  acquirer  skills?  Focusing  first  on  the  most  right-­‐skewed   attrition  pattern,  we  observe  that  the  average  adjusted  R-­‐square  jumps  from  9.08%  for  no  simulated   acquirer   skills   to   26.13%   for  𝜎!"  equal   to   5%.   This   clear   increase   reveals   a   true   reactivity   of   the   adjusted  R-­‐square  to  simulated  acquirer  skills.  This  behavior  also  is  nearly  constant  across  the  seven   simulated  attrition  patterns.  Figure  2,  Panels  A  and  B,  highlight  this  flat  behavior  of  the  adjusted  R-­‐ square.   Thus,   the   main   shortcomings   of   the   adjusted  R-­‐square   as   an   indicator   of   the   presence   of   acquirer  skills  stem  from  its  primary  function,  as  a  measure  of  the  goodness  of  fit  more  than  a  formal   statistical  test  statistic  and  its  indiscriminate  responsiveness  to  value-­‐creating  and  value-­‐destroying   skills.    

(15)

14  

The  FE  Fisher  test  is  a  formal  statistical  test  with  available  asymptotic  p-­‐values,  so  it  offers  the   most  interesting  test  for  the  presence  of  acquirer  skills.  As  highlighted  in  Section  2,  we  face  a  caveat   for  analyzing  acquirer  CAR:  the  underlying  null  hypothesis  is  that  acquirer  FE  (acquirer-­‐specific  CAR)   are  jointly  equals  to  0.  The  Fisher  test  is  therefore  designed  to  detect  the  presence  of  at  least  one   skilled  acquirer  in  the  M&A  sample  under  scrutiny,  but  it  offers  no  input  about  the  nature  of  the  skills   (value-­‐creating   or   value-­‐destroying)   or   their   frequency.   Table   3,   Panel   C,   summarizes   the   B&W   simulation,   including   the   average   values   along   the   various   attrition   patterns   and  𝜎!"  values.   With   these   average   percentages,   we   study   the   FE   test   size   when   we   simulate   no   acquirer   skills   (i.e.,   frequency   of   rejection   of   the   null   hypothesis   of   no   acquirer   skills   when   there   are   none,   or   type   I   error),  as  well  as  the  power  of  the  FE  test  (frequency  of  rejection  of  the  null  hypothesis  when  there   are  acquirer  skills,  or  1-­‐  type  II  errors).    

The  analysis  of  average  FE  Fisher  Statistic  values  reveals  promising  features:  across  all  attrition   patterns,  it  is  increasing  in  𝜎!".  That  is,  the  more  intensively  we  simulate  acquirer  skills,  the  higher   the  FE  test  values  are  on  average.  This  increase  also  is  stronger  for  the  left-­‐skewed  attrition  patterns,   which  is  a  desirable  result  because  this  sample  contains  many  more  repetitive  acquirers,  such  that   acquirer  skills  should  be  easier  to  detect.    

Turning  to  the  FE  Fisher  Statistic  size,  we  observe  that,  at  each  confidence  level  for  right-­‐skewed   attrition   patterns,   it   is   vastly   over-­‐sized.   For   example,   for   attrition   pattern   2   (most   relevant   with   respect  to  the  real  M&A  sample,  as  highlighted  in  Figure  1),  the  null  hypothesis  of  no  acquirer  skills  is   rejected  15.5%  of  the  time  at  a  10%  confidence  level,  11.30%  at  a  5%  confidence  level,  and  6.00%  at   a  1%  confidence  level.  This  size  issue  depends  on  the  attrition  pattern,  such  that  for  the  left-­‐skewed   pattern,  it  almost  disappears  (rejection  rates  fall  to  11.40%,  6.50%,  and  1.40%  at  the  10%,  5%,  and   1%   confidence   levels).   As   these   results   highlight,   the   LSDV   FE   test   is   well   suited   to   test   for   the   presence  of  at  least  one  significant  FE  when  there  are  many  repeated  units  of  observation  in  a  panel,   but  it  is  not  well  suited  when  the  sample  incorporates  many  one-­‐time  (or  a  limited  number  of  time)   units.  Yet  M&A  samples  typically  contain  such  units.  The  use  of  the  B&S  setup  to  detect  acquirer  skills   is  therefore  potentially  highly  misleading.  

Finally,   we   observe   that   the   power   of   the   FE   test   increases   with  𝜎!"  and   from   right-­‐   to   left-­‐ skewed  attrition  patterns  (as  is  clearly  observable  in  Figure  2,  Panel  B).  

 

5.  Testing  the  Presence  of  Acquirer  Skills    

(16)

15  

Understanding  the  origin  of  the  FE  Fisher  test  size  issue  is  a  first  step  towards  finding  a  solution.   Some   intuition   may   be   obtained   starting   from   the   well-­‐known   expression   that   relates   the   Fisher   statistic  to  the  R-­‐square:  

  𝐹 𝐽,𝑁−𝐾 = !!!!∗! ! !!!! !!!               (7)    

where  𝑅!  is   the  R-­‐square   of   the   unconstrained   regression,  𝑅

∗!  is   the  R-­‐square   of   the   constrained  

regression,  𝑞  is  the  number  of  restrictions,  𝑁  is  the  number  of  observations  and  𝑘  is  the  number  of   estimated  coefficients  in  the  unrestricted  model.  

Let  us  take  the  simplest  setup:  a  regression  with  only  firm  FE,  two  groups  of  firms  (of  size  𝑁!  and  

𝑁!  respectively),   the   first   group   with   firms   observed   only   once   and   the   second   group   with   firms   observed  𝑡  times.  The  regression  equation  of  the  full  model  takes  the  following  form:  

 

𝑦!,! =𝑐+𝛼!!𝐷!,!+⋯+𝛼!!! 𝐷!!,!+𝛼!!  𝛾!𝐷!!!!,!+⋯+𝛼!!! 𝐷!!!!!,!+𝜀!,!     (8)  

 

where  𝛼!!  are  the  FE  for  the  first  group  of  firms  and  𝛼!!  for  the  second.  The  null  hypothesis  is   therefore:  

 

𝐻!:  𝛼!!=⋯=𝛼!!! =𝛼!!=⋯=𝛼!!! =0         (9)  

 

  In  this  simplified  setup,  𝑁=𝑁!+ 𝑁!  ×𝑡 ,  𝑘=𝑁!+𝑁!,  𝑞= 𝑁!−1 +𝑁!,  and,  for  the   constrained  model,  𝑅∗!=0.  The  Fisher  statistic  becomes  therefore:  

 

𝐹 𝐽,𝑁−𝐾 =!!!!!!

!!!!!

(!!!!)!!!           (10)  

 

If  we  take  the  limit  when  𝑁! →𝑁  (the  sample  is  only  composed  of  firms  observed  only  one  time),  we   obtain:     lim!!→! !! !!!! !!! !! (!!!!)!!!=lim!!→! !! !!!!×!!lim! !!!!! (!!!!)!!!=  ∞  ×0       (11)    

Because   the   R-­‐square   of   a   fixed-­‐effects   regression   containing   only   one     observation   per   firm   (𝑁!=𝑁)  is  100%  and  if  𝑁!→𝑁,  𝑁!→0.  So,  in  the  limit,  the  Fisher  statistic  is  indeterminate.    

(17)

16  

  M&A  data  panel  samples  are  very  specific  in  that  they  are  characterized  by  the  presence  of   many  one-­‐time  acquirers  (41.25%  in  our  sample).  The  indeterminacy  in  the  limiting  case  of  a  sample   composed  only  of  one-­‐time  acquirers  suggests  that  this  may  be  at  the  origin  of  the  Fisher  test  size   issue.    

 

 

5.2.  Resampling  Based  Method  for  Detecting  Acquirer  Skills  (RBSD)  

If  the  presence  of  one-­‐time  acquirers  is  at  the  origin  of  the  Fisher  test  size  issue,  dropping  them   from   the   sample   is   an   easy   cure.   Starting   from   this   insight,   this   section   introduces   a   procedure   designed   to   be   as   powerful   as   possible   to   detect   acquirer   skills   if   they   are   present,   referred   hereinafter  to  RBSD.  Our  goal  is  to  make  the  number  of  acquisitions  by  acquirer  constant.    

A   first   and   obvious   solution   is   to   limit   the   sample   to   acquirers   that   completed   exactly  𝑇   acquisitions.  Then  each  acquirer  FE  can  be  estimated  using  the  same  number  of  observations,  and   the  FE  Fisher  test  builds  on  Student  laws  with  the  same  degrees  of  freedom.  There  is  a  serious  caveat   to  this  approach  though:  the  drastic  reduction  in  sample  size,  affecting  the  power  of  the  test.  In  our   sample,  only  256  acquirers  completed  exactly  5  acquisitions,  but  781  completed  5  or  more.  To  fix  this   issue,  the  RBSD  resampling  algorithm  is  as  follows:  

(i) Choose  a  given  number  of  acquisitions  𝑇  by  an  acquirer;  

(ii) Select  all  acquisitions  by  acquirers  having  completed  at  least  T  acquisitions.   (iii) Repeat  1,000  times:  

a. for   acquirers   having   completed   strictly   more   than  T   acquisitions,   random   draw   exactly  T  acquisitions  among  their  transactions;    

b. using  the  sample  of  M&A  acquisitions  selected  in  the  previous  step,  compute  the  FE   Fisher  statistic  and  test  whether  it  is  significant  against  the  Fisher  distribution  at  10%,   5%,  and  1%  confidence  levels.  

(iv) Report  the  average  FE  Fisher  Statistic  value  across  the  1,000  generated  samples  and  the   percentage  of  statistically  FE  Fisher  tests  at  10%,  5%,  and  1%  confidence  levels.  

We  replicate  the  B&W  study  that  applies  the  B&S  approach  from  Section  3  to  study  the  FE  Fisher   test  size  and  power  computed  using  our  proposed  RBSD  procedure.  We  implement  it  for  T  ranging   from  2  (minimum  possible  value  to  measure  acquirer  FE)  to  8  (reflecting  a  marginal  percentage  of   acquirers,  4.46%  in  our  sample).  To  ensure  the  comparability  of  the  results  with  the  B&W  simulation   study,  we  fixed  the  number  of  acquisitions  to  500  by  randomly  drawing  500/𝑇  acquirers  from  the   original  sample  in  step  (ii).  We  limit  this  investigation  to  the  FE  Fisher  Statistic,  because  the  R-­‐square   and  adjusted  R-­‐square  goodness  of  fit  statistics  do  not  offer  statistical  tests.    

(18)

17  

The  results  are  in  Table  4,  whose  organization  follows  that  of  Table  3,  Panel  C,  except  that  here   we  report  average  FE  Fisher  Statistic  values  and  percentages  of  statistically  significant  tests  at  10%,   5%,  and  1%  confidence  levels  for  balanced  panels  of  numbers  of  acquisitions  ranging  from  2  to  8.  In   particular,   in   Table   4,   Panel   A,   we   observe   that   average   FE   test   values   are   growing   in  𝜎!",   which   drives  simulated  acquirer  skills,  and  in  the  number  of  acquisitions.  These  are  expected  and  desirable   features.   The   FE   Fisher   test   also   checks   the   null   hypothesis   that   no   acquirer   FE   is   significantly   different   from   zero.   We   observe   in   Table   4,   Panel   A,   that   the   higher   the  𝜎!",   the   greater   the   probability  that  at  least  one  acquirer  FE  will  be  statistically  significant.  This  has  again  to  be  expected:   the  higher  the  number  of  acquisitions  by  the  acquirer,  the  lower  are  the  FE  standard  errors,9  and  the   higher  is  the  FE  Fisher  Statistic  value.    

In  the  size  analysis,  a  striking  difference  with  respect  to  the  B&S  size  (Table  3,  Panel  C)  emerges.   That   is,   the  FE  Fisher  Statistic   average   rejection   of  the  null  hypothesis   when  the   null  hypothesis   is   true  (𝜎!"  =  0%)  is  in  the  order  of  magnitude  of  the  corresponding  confidence  levels,  as  verified  for  all   numbers  of  acquisitions.  The  FE  Fisher  test  based  on  RBSD-­‐generated  samples  thus  is  correctly  sized.  

Finally,  the  improved  size  of  the  RBSD-­‐based  FE  Fisher  test  does  not  come  at  the  cost  of  a  loss  of   power.   Comparing   average   rejection   percentages   when   the   null   hypothesis   is   false   (𝜎!" >0%)   between   the   B&S   (Table   3)   and   RBSD   (Table   4)   approaches,   and   focusing   on   attrition   pattern   2   (relevant   for   the   M&A   sample,   Figure   1),   we   observe   similar   rejection   rates.   For   example,   with  

𝜎!" =3%,   the   B&S   average   rejection   rates   are   67.10%,   56.70%,   and   37.40%   at   10%,   5%,   and   1%   confidence  levels,  respectively.  The  corresponding  RBSD-­‐based  rejection  rates  are  72.90%,  61.30%,   and   38.80%   for   M&A   samples   of   four   acquisitions.   Increasing   the   number   of   acquisitions   used   to   detect   skills,   as   we   expected   intuitively,   improves   the   power   of   the   test.   Simulating   skills   with  

𝜎!" =3%  (5%  of  acquirers,  on  average,  are  imputed  a  positive  skill  𝐴𝑅  of  more  than  3%)  and  using   sequences  of  5  acquisitions,  the  average  rejection  rates  are  79.90%,  70.80%,  and  44.40%  at  10%,  5%,   and   1%   confidence   levels,   respectively.   Using   8   acquisition   sequences,   the   average   rejection   rates   jump  to  89.90%,  83.40%,  and  66.80%,  respectively.  Wooldridge  (2002,  p.  274)  emphasizes  that  “with   a   large  𝑇  (number   of   periods),   the  𝑐(𝑖)  (fixed   effects)   can   be   precise   enough   to   learn   something   about   the   distribution   of  𝑐(𝑖).   With   small  𝑇,   the  𝑐(𝑖)  can   contain   substantial   noise.”   This   is   exactly   what  we  observe  in  our  simulations.  

 

                                                                                                                         

9  This   increase   with   FE   estimation   precision   as   the   number   of   acquisitions   by   acquirer   grows   also   can   be  

observed  when  simulating  attrition  patterns.  Figure  3  displays  the  average  values  of  FE  standards  errors  along   the  seven  attrition  patterns  simulated  in  Section  3.  They  decrease  steadily  from  right-­‐skewed  to  left-­‐skewed   attritions,  as  expected.  

References

Related documents